基于随机化方法的叶绿素a含量神经网络预测模型输入因子敏感性分析

为了区分神经网络预测模型输入因子敏感性强弱,以探寻水体叶绿素a含量变化的主要影响因素,引入随机化方法,分别采用偏导、连接权值、改进连接权值、百分比扰动及改进扰动方法对叶绿素a含量神经网络预测模型输入因子进行1 000次敏感性分析,以计算结果均值对输入因子敏感性进行评价。结果表明:引入随机化方法后,敏感性分析结果稳定,研究区域pH相对敏感度最高,光照、降雨量、极大风速相对敏感度最小。受输入因子波动范围过大影响,百分比扰动方法与其他敏感性分析方法得到的结论不一致;对扰动方法进行改进,基于输入因子标准差扰动进行敏感性分析,光照、降雨量、极大风速相对敏感度分别为0.032、0.030、0.029,pH相对敏感度为0.148,因子敏感性强弱与其他方法一致;改进的扰动方法物理概念清晰,耗机时少,易实现。研究结果可为基于神经网络分析水体水华主要影响因素提供方法,为水体治理措施有效开展提供研究基础。

Sensitivity analysis on input factors of Chlorophyll-a content neural network prediction model based on randomized method

A randomization method is introduced to distinguish the sensitivity of the input factor based on the neural network prediction model and research the main influence factors of the chlorophyll-a content variation in the water. 1 000 times sensitivity analysis on the chlorophyll-a content neural network have been executed to prediction model input factors by using the partial derivative， the connection weight，the optimized connection weight，the percentage perturbation and the improved perturbation method respectively. The average value isused to evaluate the sensitivity of the input factors. The result shows that the consequences of sensitivity analysis arestable after introducing the randomization method. The relative sensitivity of pH in the research area isthe highest，while the relative sensitivity of illumination intensity， rainfall and maximum wind speed arethe lowest. The result of percentage perturbation method isinconsistent with those of other sensitivity analysis methods due to the excessive fluctuation of the input factor. After improving the perturbation method，sensitivity analysis iscarried out based on input factor standard deviation perturbation. The relative sensitivity of illumination intensity，rainfall and maximum wind speed is 0. 032，0. 030，0. 029，and pH is 0. 148，and the results areconsistent with those of other type methods. The improved perturbation method has clearer physical concept， takes less time to compute and is easier to implement. The research can provide a method for analyzing the main influencing factors of water bloom based on neural network，and a fundamental research for the effective implementation of water treatment measures.